Static IR Drop Prediction with Attention U-Net and Saliency-Based Explainability

TL;DR

This work tackles fast, accurate static IR drop prediction for large PDNs under limited real-design data. It introduces AttUNet, a U-Net variant with preprocessing and vector concatenation-based attention gates to handle multi-image inputs and sparse IR-drop maps, trained via a two-phase pretrain-finetune regimen augmented with data transformations. A saliency-map-based explainability method is developed to attribute high-drop pixels to specific input features, enabling targeted PDN optimizations. Experimental results show AttUNet surpasses prior models and the 2023 ICCAD winner in real-design tests, achieving notable reductions in MAE and improvements in F1, with fast inference and meaningful guidance for design improvements.

Abstract

There has been significant recent progress to reduce the computational effort of static IR drop analysis using neural networks, and modeling as an image-to-image translation task. A crucial issue is the lack of sufficient data from real industry designs to train these networks. Additionally, there is no methodology to explain a high-drop pixel in a predicted IR drop image to its specific root-causes. In this work, we first propose a U-Net neural network model with attention gates which is specifically tailored to achieve fast and accurate image-based static IR drop prediction. Attention gates allow selective emphasis on relevant parts of the input data without supervision which is desired because of the often sparse nature of the IR drop map. We propose a two-phase training process which utilizes a mix of artificially-generated data and a limited number of points from real designs. The results are, on-average, 18% (53%) better in MAE and 14% (113%) in F1 score compared to the winner of the ICCAD 2023 contest (and U-Net only) when tested on real designs. Second, we propose a fast method using saliency maps which can explain a predicted IR drop in terms of specific input pixels contributing the most to a drop. In our experiments, we show the number of high IR drop pixels can be reduced on-average by 18% by mimicking upsize of a tiny portion of PDN's resistive edges.

Figures (8)

• Figure 1: Components of our methodology: (a) image translation; (b) data augmentation and two-step training; (c) inference with attention U-Net; (d) saliency map generation; (e) explaintion and diagnosis of predicted hotspots.
• Figure 2: Block diagram of AttUNet. Additional embedded components compared to a U-Net structure are highlighted.
• Figure 3: Vector concatenation-based attention gate.
• Figure 4: Augmenting the training data via five transformations applied to each image-based input.
• Figure 5: Illustration of the cosine annealing learning rate.